Open-source Raven agent runs 42 autonomous rounds over 4 days to ship a playable FPS
lxfater · x · 2026-09-29
The open-source multi-agent system Raven demonstrates long-horizon autonomy: given only a requirements doc, it ran 42 rounds of plan→develop→review over about 4 days and built a playable Godot 4 FPS boss fight.
Key points:
- Consistent iteration: the boss evolved through 5 stages, from a cylinder with an orange light orb into a four-legged beast with a molten core and five attack patterns.
- Orchestrator design: Raven is a "harness of harnesses" that can dispatch work to external agents like Claude Code and Codex.
- Project: EverMind-AI/Raven on GitHub (4.7k stars), with demo video, posters and site in the README.
A useful case study for anyone building long-running autonomous agents.
More from coding & agent
- build-eval + hillclimb bring real eval tooling to Claude Code agents — Arindam_1729 · 2026-09-29
- One prompt turns Claude Opus 5.5 into a professional motion design studio — socialwithaayan · 2026-09-29
- 15% of all skills are agent workflows, stats breakdown shows — zainhas · 2026-09-29
- 180k typed tool-calling decisions released so tiny models can route tools — MaziyarPanahi · 2026-09-29
- Who grades the grader? A practical method for testing checks written by coding agents — jonah_omninode · 2026-09-29
- Perplexity launches multi-agent trading, running models like GPT-6 Astra and Grok 4.7 on Coinbase — AravSrinivas · 2026-09-29